{"id":"de1480cf-0e88-4b46-a6cb-8052a5c970b8","arxiv_id":"2411.15271","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"EADReg combines BGMM outlier removal with an efficient autoregressive diffusion correspondence generator, achieving SOTA outdoor LiDAR point cloud registration at near-convolutional speed.","lead":"This paper proposes EADReg, a two-stage registration system that first removes outlier points with a bidirectional Gaussian mixture model, then refines alignment with an autoregressive diffusion model that generates only local correspondences. It reports state-of-the-art accuracy on three outdoor LiDAR benchmarks (KITTI, NuScenes, Apollo-SouthBay) with runtime close to fast convolutional methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Autoregressive diffusion loop is internally inconsistent: the warping step changes the KNN candidate set and the diffusion state, but the training loss (Eq. 20) does not simulate this loop, so the core novelty is unverified.","rationale":"The reader identified the top-K locality assumption as the weakest assumption, and that is a real limitation. However, I find a more load-bearing internal-consistency issue: the autoregressive diffusion inference, which is the paper's claimed novelty, is not a well-defined diffusion process as described. The conditioning features F^t_G are recomputed from a warped source at every denoising step, while the diffusion state C_t is a noisy correspondence over the previous candidate set. The paper provides no training objective or algorithmic description showing that the model learns to handle this changing candidate set, and the loss in Eq. 20 is a single-step MSE that does not explicitly include the coarse initialization or the sequential warping. The Softmax at inference (Eq. 12) versus raw MSE in training (Eq. 20) is a second, concrete train-test mismatch. These issues do not prove the empirical numbers are wrong, but they mean the central claim—that a diffusion-based correspondence generator achieves SOTA accuracy and speed—is not fully supported by the text as written. The paper should be accepted only if the authors clarify and validate the autoregressive training procedure, for example by simulating the inference loop during training and ablating the warping steps. This is consistent with the reader's CONDITIONAL verdict, so I recommend no change to the verdict.","tokens_in":21069,"tokens_out":7293,"duration_ms":69156,"concrete_test":"Run the released code with the autoregressive warping loop disabled: keep the diffusion steps and the coarse-warped features F^T_G fixed for every t, skipping Eq. 14 and line 8 of Algorithm 1, and also remove the Softmax in Eq. 12 to match Eq. 20. If KITTI RTE/RRE change by less than the reported 0.7 cm/0.03 degree margin over HRegNet, the autoregressive diffusion mechanism is not the source of the gains; if they change substantially, retrain with a loss that explicitly simulates the inference-time warping loop and verify the numbers are preserved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The fine-stage diffusion model is the paper's core novelty, but as written its training and inference do not match. At inference (Algorithm 1), each denoising step produces a transformation from the predicted correspondence (Eq. 13), warps the source points (Eq. 14), and recomputes the geometric features F^t_G (Eq. 10), including a new top-K candidate set. The diffusion state C_t, however, is a noisy correspondence over the previous candidate set; feeding it as input alongside new candidates has no defined meaning, and the DDIM update (lines 9-10) mixes a cleaned correspondence over old candidates into a new noisy state over different candidates. This is not a valid reverse diffusion process. The training loss (Eq. 20) is a single-step MSE between F_f[C_t,t,F^t_G,F_D] and C_gt, where C_gt is built from GT-warped source points; the paper never states that the training loop simulates the sequential warping, the coarse initialization, or the changing candidate sets used at inference. Additionally, Eq. 12 applies Softmax to the network output at inference, while Eq. 20 minimizes raw MSE without Softmax, a further train-test mismatch. If the autoregressive warping is not trained, the reported 0.040 m RTE may come from repeated SVD refinement rather than from the learned diffusion process, and the probabilistic autoregressive claim (Eq. 16) is not substantiated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes EADReg, a coarse-to-fine pipeline for outdoor LiDAR point cloud registration. In the coarse stage, a Bi-directional Gaussian Mixture Model (BGMM) module rejects outlier GMM components and their attached points, after which a lightweight network computes coarse correspondences and an SVD-based transformation. In the fine stage, a diffusion model denoises a top-K correspondence matrix, and an autoregressive-style inference loop warps the source cloud by the predicted transformation at each denoising step before recomputing the candidate set. The method is evaluated on KITTI, NuScenes, and Apollo-SouthBay, reporting state-of-the-art RTE/RRE and runtime comparable to convolutional baselines (e.g., 0.040 m RTE and 0.117 degrees RRE on KITTI at 129.4 ms), with ablations on the number of GMM clusters, diffusion steps, candidate size K, and rotation loss weight.","tokens_in":21375,"tokens_out":5073,"duration_ms":48316,"significance":"If the results hold, EADReg would be a meaningful advance: it demonstrates that a diffusion-based correspondence generator can be accurate and fast for large-scale outdoor LiDAR registration, and its top-K formulation avoids the prohibitive memory cost of dense N_S x N_T correspondence prediction. The BGMM outlier removal is training-free and introduces no extra learnable parameters, and the experiments cover three standard benchmarks with consistent baselines. However, the significance of the specific diffusion contribution is currently not established because the training objective in Eq. (20) does not simulate the autoregressive loop used at inference, and the inference-time softmax in Eq. (12) is not matched by the training loss. The manuscript also omits key implementation details (GMM fitting procedure, detector-descriptor backbone) needed to reproduce or even fully evaluate the method.","major_comments":[{"comment":"The paper's core novelty, the autoregressive diffusion loop, is not trained as described. In Algorithm 1, lines 4-10, each denoising step computes a transformation from the denoised correspondence, warps the source points (line 8), recomputes the geometric features F^t_G with a new KNN candidate set, and then forms C_{t-1} via the DDIM update over the previous candidate set. The training loss in Eq. (20), however, is a single MSE between F_f[C_t, t, F^t_G, F_D] and C_gt, where C_gt is built once from GT-warped source points; the paper never states that training simulates the sequential warping, the coarse initialization, or the changing candidate sets used at inference. As written, the DDIM update mixes a cleaned correspondence over old candidates into a new noisy state whose features come from different candidates, which is not a valid reverse diffusion process. This mismatch must be resolved (e.g., by training with the inference loop, or by explicitly reformulating the method as iterative SVD refinement rather than autoregressive diffusion) before the probabilistic autoregressive claim in Eq. (16) can be accepted.","section":"Sec. 3.3 (Eq. 20) and Algorithm 1"},{"comment":"The inference procedure applies Softmax to the network output (Eq. 12) to obtain candidate weights that are then used in the weighted SVD of Eq. (13), but the training loss in Eq. (20) minimizes raw MSE between the network output and C_gt without any Softmax. These are not equivalent objectives for the downstream transformation estimate, so the training signal does not directly supervise the quantity used at inference. The authors should align the training and inference operations (for example, supervise the softmaxed weights with a cross-entropy or KL term, or remove Softmax from inference) and rerun the experiments.","section":"Sec. 3.3 (Eq. 12) vs. Eq. (20)"},{"comment":"The BGMM module is under-specified. Eq. (1) defines a GMM, but the paper does not state how the J components are fit (EM initialization, number of iterations, covariance regularization), how Eq. (2) is computed in practice (the notation N_k(Theta_T_j) and N_1(Theta_S_i) is ambiguous), or how outlier GMM components are mapped back to individual points. Since BGMM is a central contribution and its removal is ablated in Table 3, these details are necessary for reproducibility and for assessing whether the coarse-stage gains come from the proposed filtering rule rather than from the GMM fitting configuration.","section":"Sec. 3.2 (Eqs. 1-2)"},{"comment":"The fine stage assumes that after coarse warping the true target of each source point lies among its K nearest target neighbors. This locality premise is load-bearing for both accuracy and efficiency, but it is validated only on the three benchmark datasets with interval-10 pairs. The paper should report results under larger initial drift, partial overlap, or reduced K to show the scope of the assumption and to characterize the failure mode of the diffusion refinement; as written, the robustness claim to 'perturbations in outdoor LiDAR point clouds' is broader than the evidence supports.","section":"Sec. 3.3 (Eq. 10) and Table 6"}],"minor_comments":[{"comment":"The title of Algorithm 1 and the phrase in Sec. 3.3 misspell 'autoregressive' as 'autogressive'; please correct.","section":"Algorithm 1 and Sec. 3.3"},{"comment":"The table header uses 'K' for the number of sampling steps, which conflicts with the correspondence-candidate size K in Table 6 and Eq. (10); it should be labeled 'S'.","section":"Table 7"},{"comment":"The text says 'the average RRE of the coarse stage with 0.119cm is better than the learning based method DGR', but Table 3 reports Only Coarse as RTE 0.119 m and RRE 0.373 degrees; the metric names and units are inconsistent.","section":"Sec. 4.3, Network Structure paragraph"},{"comment":"Eq. (18) does not define how tilde t_l relates to the coarse and fine stage transformations, and the index l in {c,f} is not made explicit in the definitions of tilde R_l and tilde t_l; please specify which transformations enter the loss.","section":"Eq. (18)"},{"comment":"The reference list appears twice in the submission; please remove the duplicate.","section":"References"},{"comment":"The sentence 'the influence of different sampling steps will be discussed in the Sec.' lacks a section number; please provide the correct cross-reference.","section":"Sec. 4.1"},{"comment":"The notation p(C^T, ..., C^t) = prod_t p(C^t | C^T, ..., C^{t-1}) is inconsistent with the reverse-time indexing in Algorithm 1 and with the conditioning on the features F^t_G; please align the notation with the actual inference procedure.","section":"Eq. (16)"}],"recommendation":"major_revision","confidential_remarks":"The main concern is the train-test mismatch in the autoregressive diffusion loop: as written, the paper does not demonstrate that the diffusion model is trained for the inference procedure it claims, so the state-of-the-art numbers may be driven by repeated SVD refinement rather than by learned denoising. I would ask the authors to either retrain with the inference loop or substantially rewrite the method description and re-evaluate. The paper could become publishable if this is resolved and the missing implementation details are provided, but the current form is not sufficiently verified."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper proposes EADReg, a coarse-to-fine LiDAR registration framework with a BGMM outlier-rejection module and a diffusion-based correspondence refinement stage. What is genuinely new is the combination: using GMMs to purify point clouds before a top-K diffusion correspondence generator, and framing the reverse process as autoregressive warping. The experiments are thorough and internally consistent, with ablations over J, K, S, and alpha, and the reported runtime (129 ms on KITTI) is competitive with convolutional methods. The gains over strong baselines like HRegNet are real but modest (0.7 cm RTE on KITTI, 2 cm on NuScenes). Credit where due: the BGMM idea is simple and apparently effective, and the top-K formulation is a sensible answer to the point-scale problem in outdoor LiDAR. The soft spots are not minor. The core novelty, autoregressive diffusion inference, is not actually trained. The loss (Eq. 20) is a single-step MSE between the network output and the GT correspondence, with no simulation of the sequential warping, the changing KNN candidate sets, or the transformation composition used in Algorithm 1. At inference, each step recomputes the candidate set and geometric features after warping, but the noisy diffusion state C_t still refers to the previous candidate set. That is an undefined input combination. The softmax in Eq. 12 vs. the raw MSE in Eq. 20 adds another mismatch. So the paper's central probabilistic-autoregressive claim is unsupported. We cannot tell whether the accuracy comes from the diffusion process or simply from repeated SVD refinement on a fixed model. The paper also omits GMM fitting details and the backbone architecture, defers to a missing supplementary, and provides no code. Given these gaps, the reported numbers are not independently checkable. I would still send this to peer review. The problem is interesting, the experiments are substantive, and the BGMM module plus top-K diffusion design could be a useful subfield contribution after a major revision. What I want in revision is either a training loss that includes the autoregressive loop (with teacher forcing or scheduled sampling) or an explicit statement that the warping is a test-time heuristic, with the architecture and code released. If the authors cannot clarify the discrepancy, the paper's contribution shrinks to an engineering artifact. For a reader: if you work on LiDAR registration or diffusion-based matching, this is worth a skim for the BGMM idea, but do not build on the autoregressive claim until the code is out. I would bring it to a reading group only if we want to discuss train-test mismatches in generative model inference. My honest verdict: reject in current form, but with a request for major revision rather than a desk reject.","headline":"The autoregressive diffusion loop is the paper's centerpiece, but its training objective does not match the inference procedure, leaving the main novelty unverified.","tokens_in":749,"tokens_out":975,"would_cite":false,"duration_ms":37804,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"EADReg claims a diffusion-based correspondence generator can register outdoor LiDAR scans at centimeter accuracy while running as fast as convolutional methods.","keywords":["point cloud registration","LiDAR","diffusion models","autoregressive diffusion","Gaussian mixture model","outlier rejection","coarse-to-fine registration","outdoor scenes"],"falsifier":"Take the trained EADReg on KITTI and, after the coarse stage, add a known translation offset of 5-10 m to the warped source clouds before the fine stage. If the method's recall or RTE degrades sharply, the top-K locality assumption is the load-bearing premise; if it stays flat, the diffusion stage is recovering correspondences beyond the coarse neighborhood. A second check: sweep K from 3 to 32 on a held-out set and see whether RTE drops materially, which would contradict the paper's ablation conclusion that K barely matters.","tokens_in":20880,"feed_emoji":"🚗","tokens_out":8062,"duration_ms":70152,"temperature":0.7,"pith_summary":"EADReg claims that diffusion models, usually seen as too slow and memory-hungry for large outdoor scans, can be made the basis of a fast and accurate LiDAR registration pipeline. The method combines a training-free outlier-rejection step, bidirectional Gaussian mixture models, with a diffusion process that generates correspondences only among the top-K nearest neighbors of each source point rather than across the whole target cloud. Inference is treated autoregressively: each denoising step produces a small transformation that warps the source points for the next step, and the final pose is the product of the accumulated transformations. On KITTI, NuScenes, and Apollo, the paper reports centimeter-level translation errors (0.040 m on KITTI, 0.090 m on NuScenes) with runtime around 129 ms and 112 ms, comparable to fast convolutional baselines. If true, this shows that generative correspondence estimation need not be sacrificed for speed in outdoor robotics.","feed_headline":"Diffusion model hits 4 cm accuracy on outdoor LiDAR registration","feed_subtitle":"EADReg pairs GMM outlier rejection with fast autoregressive diffusion, keeping generative registration practical for driving.","key_machinery":"Three mechanisms carry the argument. First, the Bi-directional Gaussian Mixture Model (BGMM) outlier-removal module fits a GMM to each scan and removes points attached to Gaussian components whose top-K counterpart in the other scan is not reciprocal, purifying the clouds at no extra training cost. Second, the restricted diffusion correspondence generator predicts, for each source point, a distribution over its K nearest target neighbors ($\\hat{C} \\in \\mathbb{R}^{N_S \\times K}$) conditioned on geometric features, descriptor features, noise level, and time step, instead of a global $N_S \\times N_T$ correspondence matrix. Third, the autoregressive inference loop converts each denoising step into a rigid transform that warps the source points for the next step, with the final pose formed by the product of transforms. Together, the coarse purification makes the top-K locality assumption valid, and the locality assumption is what collapses the cost of diffusion-based correspondence generation.","core_discovery":"The paper's claim is that point-to-point correspondence generation for outdoor LiDAR can be reformulated as a sequence of small, local denoising problems instead of one dense global matching problem. The coarse stage fits a Gaussian mixture model to each scan, removes components whose mutual nearest neighbor in the other scan is not among their top-K, and uses the purified points to estimate an initial transform by weighted SVD. The fine stage then trains a lightweight conditional diffusion model to denoise a correspondence matrix $\\hat{C} \\in \\mathbb{R}^{N_S \\times K}$ that assigns each source point a distribution over its K nearest target candidates. During inference, each denoising step yields a transform that warps the source cloud for the next step, so the sequence of transforms accumulates into the final registration. The paper argues this design is what lets a diffusion-based method reach 0.040 m RTE and 0.117 degrees RRE on KITTI with 129.4 ms runtime, beat prior diffusion PCR methods on outdoor data, and produce the only centimeter-level RTE on NuScenes among the compared methods.","pith_inferences":["An untested consequence is that the same coarse-purify-then-local-diffusion recipe could transfer to other sparse 3D sensors, such as solid-state LiDAR or radar, if their point density keeps the top-K assumption intact.","The autoregressive inference loop is stated only for rigid-body transforms; the history-buffer mechanism is a generic sequence-modeling scaffold that could in principle be applied to deformable or non-rigid matching, though the paper does not explore this.","The paper validates K through ablations on KITTI; a stronger test would be to recompute those ablations under larger frame intervals or added synthetic drift, where the number of nearest neighbors needed to enclose the true match would grow."],"forward_implications":["Diffusion-based correspondence generation can be deployed in real-time outdoor pipelines; the paper reports 129.4 ms per pair on KITTI and 112.4 ms on NuScenes.","The BGMM outlier-removal step improves coarse registration without introducing trainable parameters; removing it raises RTE from 0.040 m to 0.042 m in the paper's ablation.","Because GPU memory grows steeply with K (from 19,083 MB at K=1 to 23,571 MB at K=7 on KITTI), global dense correspondence prediction is impractical for large scans, so the top-K formulation is necessary for training feasibility.","The diffusion step count can be kept small: performance saturates around three sampling steps, so DDIM acceleration does not cost accuracy.","The coarse-to-fine autoregressive design makes the diffusion model behave as a refinement module, so its robustness gains come without the inference cost usually attributed to generative registration methods."],"supporting_citations":[{"why":"Supplies the hierarchical detector-descriptor backbone EADReg reuses and the strongest convolutional baseline it must beat on all three datasets.","marker":"[39]"},{"why":"The prior diffusion PCR method that directly regresses transformation parameters; EADReg positions its autoregressive top-K correspondence design against it.","marker":"[7]"},{"why":"The diffusion correspondence-search baseline that generates global dense correspondences; EADReg contrasts its local top-K formulation with this costly approach.","marker":"[58]"},{"why":"Defines the denoising diffusion probabilistic model whose forward and reverse processes EADReg adapts for correspondence generation.","marker":"[19]"},{"why":"DDIM sampling is used during testing to cut the number of diffusion steps while preserving accuracy.","marker":"[49]"},{"why":"KITTI odometry data provide the primary benchmark and the reported 0.040 m RTE / 0.117 degree results.","marker":"[15]"},{"why":"NuScenes provides the second benchmark where EADReg reports the only centimeter-level RTE among compared methods.","marker":"[3]"},{"why":"A GMM-based outdoor registration method that motivates using Gaussian mixture modeling for LiDAR; EADReg instead uses GMMs for outlier rejection.","marker":"[46]"},{"why":"GMM-based registration baseline whose modeling approach is repurposed; EADReg contrasts its own bidirectional outlier removal with direct GMM alignment.","marker":"[64]"}],"fun_headline_variants":["Autoregressive diffusion sharpens LiDAR registration to 4 cm","Diffusion gets LiDAR registration down to 4 cm in 129 ms","EADReg: autoregressive diffusion for centimeter-accurate LiDAR pose","Autoregressive diffusion reformulates dense matching to 4 cm LiDAR registration","Diffusion-based PCR goes outdoor: 4 cm accuracy, 129 ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that after the coarse stage, the true match for every source point is among its K nearest neighbors in the target scan; if the coarse transform drifts or occlusion removes the correct candidate, the diffusion stage has no way to recover it.","fun_headline_variants_meta":{"raw":{"variants":["Autoregressive diffusion sharpens LiDAR registration to 4 cm","Diffusion gets LiDAR registration down to 4 cm in 129 ms","EADReg: autoregressive diffusion for centimeter-accurate LiDAR pose","Autoregressive diffusion reformulates dense matching to 4 cm LiDAR registration","Diffusion-based PCR goes outdoor: 4 cm accuracy, 129 ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001559,"raw_usage":{"total_tokens":6263,"prompt_tokens":1017,"completion_tokens":5246,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":633,"completion_tokens_details":{"reasoning_tokens":5147}},"tokens_in":633,"tokens_out":5246,"duration_ms":36584,"temperature":1.0,"reasoning_tokens":5147,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:43:52.486719+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the trained EADReg on KITTI and, after the coarse stage, add a known translation offset of 5-10 m to the warped source clouds before the fine stage. If the method's recall or RTE degrades sharply, the top-K locality assumption is the load-bearing premise; if it stays flat, the diffusion stage is recovering correspondences beyond the coarse neighborhood. A second check: sweep K from 3 to 32 on a held-out set and see whether RTE drops materially, which would contradict the paper's ablation conclusion that K barely matters.","supporting_citations":[{"cited_title":"Hregnet: A hierarchical network for efficient and accurate outdoor lidar point cloud registration","cited_arxiv_id":null,"evidence_quote":"Supplies the hierarchical detector-descriptor backbone EADReg reuses and the strongest convolutional baseline it must beat on all three datasets."},{"cited_title":"Denois- ing diffusion implicit models","cited_arxiv_id":null,"evidence_quote":"DDIM sampling is used during testing to cut the number of diffusion steps while preserving accuracy."},{"cited_title":"nuscenes: A multi- modal dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"NuScenes provides the second benchmark where EADReg reports the only centimeter-level RTE among compared methods."},{"cited_title":"G3reg: Pyramid graph-based global registra- tion using gaussian ellipsoid model","cited_arxiv_id":null,"evidence_quote":"A GMM-based outdoor registration method that motivates using Gaussian mixture modeling for LiDAR; EADReg instead uses GMMs for outlier rejection."},{"cited_title":"Deepgmr: Learning la- tent gaussian mixture models for registration","cited_arxiv_id":null,"evidence_quote":"GMM-based registration baseline whose modeling approach is repurposed; EADReg contrasts its own bidirectional outlier removal with direct GMM alignment."}],"review_version":1}